Statistical modeling in measurement-based resource management

With the growing demand for high-quality real-time and interactive applications, there is now a call for service differentiation within the Internet. As a result, resource management must become more sophisticated to properly control the network. An important challenge is to manage the network to maximize the utilization of resources and to meet the performance requirements of admitted connections. In this dissertation, we investigate the role of statistical modeling and measurement for monitoring and controlling QoS to provide differentiated services to users.; Statistical modeling is an important tool for characterizing the behavior of the network and for providing statistical guarantees of performance. Although it is possible to allocate resources using worst-case provisioning, this can lead to underutilization. Furthermore, most applications do not require stringent performance guarantees and can withstand minor QoS infractions. Statistical provisioning allows the network to increase network utilization while bounding the probability of QoS violation. However, traffic models can be quite complex, and so measurement-based methods have been proposed for evaluating statistical performance guarantees using measurements of actual traffic.; We use an innovative approach for estimating certain QoS performance metrics, such as loss, rate, and delay, from network measurement. First, for real-time traffic in FCFS queuing systems, the performance objective is to control the loss probability through admission control. Since losses are rare, we estimate the loss probability by estimating the buffer occupancy distribution. A key advantage of our approach is the use of model selection methods to improve the robustness of such estimators. Second, for class-based queuing systems, the objective is to allocate resources to individual traffic classes to control rate and delay. We use a measurement-based monitoring and admission control algorithm that provides a minimum guaranteed rate and maximum scheduling latency—called a rate-latency bound—to traffic classes, but allows for infrequent QoS violations. Third, for Web-based applications using TCP, it is important to make sure that the network is sufficiently provisioned to provide quality in the absence of admission control. We model TCP performance at the transaction level and use this model to investigate the impact of capacity allocation on quality.